資料與限制
每個 30 公尺像元的樹冠比例,由 Landsat 4、5、7、8、9 衛星影像(USGS Collection 2 Level-2,Tier 1)逐年估計。畫面上的樹是示意:數量與樹冠比例成正比,不代表實際的單株樹木;樹種依海拔帶與都市環境隨機配置,樹葉由程式繪製;濃密森林的樹下另畫示意的林下灌木。鏡頭附近畫出細緻的立體樹,越遠越簡化;約 3 公里外只在每 2、4、8 個像元放一棵、樹冠相應加寬(覆蓋比例不變),因此立體樹可延伸到 6–14 公里,最外圈逐漸稀疏、融入地面的樹冠紋理與約 13 公尺的樹冠高度(同樣依樹冠比例)。電腦較慢時,遠處會更早開始稀疏(但樹冠在畫面上不會大於約 10 像素),並縮小範圍,以維持流暢。
怎麼算的
- 每年、每個像元取一景真實觀測(medoid 合成),因此每個數值都能追溯到單一原始影像與像元。點選地面即可查看來源影像編號。
- 像元位置保留 Landsat 原始格網(UTM,像元中心為 30 公尺整數倍),未經重新取樣。
- 樹冠比例模型以 ESA WorldCover 2021 的樹木類別為參照,用 2021 年影像訓練;在保留區塊上與參照的一致性 R² ≈ 0.71,這是與參照的一致程度,不是實地驗證的準確度。
- Landsat 5 先以臺灣重疊年份擬合的係數校正到 Landsat 7,再以 Roy 等人(2016)係數調和至 Landsat 8;模型只用 Landsat 8 影像訓練。
- 逐年正規化(全臺資料):穩定郁閉森林與長期無樹的空曠地,每年應該看起來一樣。每個圖塊、每一年,都把這兩類像元的反射率中位數對齊到該圖塊 2021 年(模型訓練年)的數值,再套用到整個圖塊後才計算樹冠比例。澎湖、金門、馬祖附近沒有足夠的參照像元,未做正規化。
限制
- 1982、1983、1985 年臺北沒有可用影像;1984、1986 年只有 3 景,屬低信心年份(時間軸以橘色標示)。
- 原始影像的定位誤差約 12 公尺(RMSE),不同年份之間可能有不到一個像元的偏移。
- 感測器差異校正(第二版):同年、同像元比對顯示 Landsat 5(TM)比 Landsat 7 低約 4.9 個百分點;第二版加入以臺灣北部重疊年份擬合的 TM→ETM+ 校正後,差距降到約 −0.4。1999 年前後的跳階由 +6,470 公頃縮小為 +2,660 公頃。
- 影像逐年漂移(已正規化):未正規化時,穩定森林的數值從 1987–98 年約 80% 逐步升到 2022–26 年約 91%(北部 2013 年後最明顯),這是影像光譜(霧霾、季節組成)逐年變化被模型讀成樹冠增加,不是新增樹冠;全臺陸地樹冠比例也因此從 65.2% 升到 72.4%,2022 年單年更達 75.3%。正規化後為 66.7% → 69.3%,2022 年與 2021 年相當(70.5%/70.3%),2016–2025 年的趨勢與獨立的 Sentinel-2 樹木機率(Dynamic World)一致。仍有的限制:1987–98 年的穩定森林仍比 2014 年後低約 4 個百分點;正規化也抹平了部分真實的區域性年際起伏,因此單一年份、幾個百分點的區域變化不宜解讀為真實增減。可開啟未正規化的版本對照。
- 模型只用 2021 年訓練,套用到 1980–90 年代時假設光譜關係不變,尚未以有日期的高解析影像獨立驗證。
- 30 公尺像元約 0.09 公頃,多數行道樹小於一個像元;樹冠比例反映的是像元內的樹冠占比。
- 某年沒有清晰觀測的像元,畫面沿用前一次觀測並以灰色條紋標示,不會被當成樹冠變化。
- 地形高度來自 Copernicus DEM GLO-30,僅供顯示,並放大 1.5 倍。
- 天氣、雲、風、雨與日夜只是畫面效果,不是實際氣象資料;雲影也不影響樹冠數值。
- 建築物是示意:某像元在 GISA 資料中第一次成為不透水面(建成)的那一年,建築才會長出來;數量依 GHSL 2018 的建成面積比例,高度依 GHSL 2018 的建物高度,位置與形狀是程式產生的,不是真實建物輪廓。GISA 只記錄「第一次建成」,所以拆除、改建或都市綠化不會顯示;2021 年以後新建的也還沒有。夜間燈光同樣只是示意。
- 地址搜尋使用 OpenStreetMap 的 Nominatim;臺灣門牌資料不完整,找不到時會逐步退到巷弄、路段或行政區,並顯示達到的精度。
- 廠區故事(含
?story= 的連結)是把同一套樹冠比例,平均到 OpenStreetMap 的廠區範圍與其周邊環帶;文中的興建日期來自所列的新聞與公司資料。廠房是 OpenStreetMap 的實際建物輪廓(不是示意),高度多為估計,出現的年份依 GISA 第一次建成的年份,或廠區樹冠消失的年份推估;半透明的是 OpenStreetMap 標示為興建中的建物。開啟故事時,廠區內不畫示意建築。
- 營收估計與獨角獸:台積電只公布全公司營收,不公布各廠營收。故事中的廠區營收是估計:每月公布的營收,依當季各製程占晶圓營收的比例,再依本廠區在各製程產能中的估計占比拆分(封測廠依先進封裝營收占比與各封測廠的估計占比),以年均匯率換成美元;範圍依各項拆分的把握程度(約 ±10–45%)。2015 年後台積電不再公布各廠產能,之後的拆分多依新聞與分析師估計。每累積 10 億美元,就有一隻獨角獸從廠房飛出(純屬趣味);「每隻 🦄 約幾棵樹」以消失的樹冠面積、每棵樹約 50 平方公尺樹冠粗估。資料與來源:
pipeline/analysis/tsmc_revenue_inputs.json。
來源
- Landsat Collection 2 Level-2 courtesy of the U.S. Geological Survey(經 Google Earth Engine 處理)
- ESA WorldCover 2021 v200,CC BY 4.0
- 建成年份:GISA 1972–2021(Ren, Huang, Yang & Zhou 2025, ISPRS J. Photogramm. Remote Sens. 220),CC BY 4.0
- 建物高度與建成面積:JRC GHSL P2023A(GHS-BUILT-H 2018、GHS-BUILT-S 2018 10 m),European Commission, Joint Research Centre
- Copernicus DEM GLO-30 © DLR e.V. 2010–2014 and © Airbus Defence and Space GmbH 2014–2018, provided under COPERNICUS by the European Union and ESA
- 地址搜尋與廠區範圍:Nominatim / © OpenStreetMap contributors(ODbL)
資料檔(Apache Arrow)每個像元都帶有 UTM 座標與來源影像索引,可用 Python(polars、pyarrow、DuckDB)直接開啟。
Data & limits
Tree cover (the share of each 30 m pixel covered by tree canopy) is estimated year by year from Landsat 4, 5, 7, 8 and 9 imagery (USGS Collection 2 Level-2, Tier 1). The trees on screen are illustrative: their number follows the tree cover, not real individual trees. Species are assigned at random by elevation band and urban setting, the foliage is drawn by code, and dense forest gets an illustrative shrub layer. Near the camera the 3D trees are detailed and they get simpler with distance. Beyond about 3 km only every 2nd, 4th or 8th pixel carries trees, with correspondingly wider crowns (the same cover), so 3D trees reach 6–14 km; the outermost band thins out into the canopy texture drawn on the ground with a canopy height of about 13 m (also following tree cover). On slower computers the thinning starts nearer (widened crowns never exceed about 10 screen pixels) and the range shrinks, to keep things smooth.
How it is computed
- For each year and pixel, one real observation is kept (a medoid composite), so every value can be traced to a single original scene and pixel. Click the ground to see the source scene ID.
- Pixels stay on the native Landsat grid (UTM, pixel centres at multiples of 30 m) and are never resampled.
- The tree-cover model uses the tree class of ESA WorldCover 2021 as its reference and is trained on 2021 imagery. On held-out blocks it agrees with that reference at R² ≈ 0.71. That measures agreement with the reference, not accuracy checked on the ground.
- Landsat 5 is first calibrated to Landsat 7 with coefficients fitted on overlapping years in Taiwan, then harmonised to Landsat 8 with the Roy et al. (2016) coefficients; the model is trained on Landsat 8 only.
- Yearly normalisation (island data): stable closed forest and long-term treeless open land should look the same every year. For each tile and year, the median reflectance of these two kinds of pixel is aligned to the tile's 2021 values (the model's training year), and that correction is applied to the whole tile before tree cover is computed. Penghu, Kinmen and Matsu do not have enough reference pixels and are not normalised.
Limits
- There is no usable imagery over Taipei for 1982, 1983 or 1985; 1984 and 1986 have only 3 scenes and are low-confidence years (marked orange on the timeline).
- The raw imagery's geolocation error is about 12 m (RMSE), so features can shift by less than a pixel between years.
- Sensor calibration (version 2): same-year, same-pixel comparisons showed Landsat 5 (TM) reading about 4.9 points lower than Landsat 7. Version 2 adds a TM→ETM+ correction fitted on overlapping years in northern Taiwan, which brings the gap to about −0.4. The step around 1999 shrank from +6,470 ha to +2,660 ha.
- Image drift between years (normalised): without normalisation, stable forest rises from about 80% in 1987–98 to about 91% in 2022–26 (most visibly in the north after 2013). That is the model reading year-to-year changes in the imagery (haze, the mix of seasons) as more canopy, not new trees. It also pushes island land tree cover from 65.2% to 72.4%, and to 75.3% in 2022 alone. After normalisation it is 66.7% → 69.3%; 2022 matches 2021 (70.5% / 70.3%), and the 2016–2025 trend agrees with the independent Sentinel-2 tree probability (Dynamic World). Remaining limits: stable forest still reads about 4 points lower in 1987–98 than after 2014, and the normalisation also flattens some real regional year-to-year variation, so regional changes of a few points in a single year should not be read as real gains or losses. You can open the non-normalised version to compare.
- The model is trained on 2021 only. Applying it to the 1980s–90s assumes the spectral relationships have not changed; it has not been independently validated against dated high-resolution imagery.
- A 30 m pixel is about 0.09 ha, so most street trees are smaller than a pixel; tree cover is the canopy share inside each pixel.
- Where a pixel has no clear observation in a year, the previous observation is shown with grey stripes; it is never counted as a change in tree cover.
- Terrain heights come from the Copernicus DEM GLO-30, are for display only, and are exaggerated 1.5×.
- Weather, clouds, wind, rain and day/night are visual effects only, not weather data; cloud shadows do not affect tree-cover values.
- Buildings are illustrative: buildings rise in the year a pixel first becomes impervious (built) in the GISA data. Their number follows the GHSL 2018 built-up share and their height the GHSL 2018 building height; positions and shapes are generated, not real footprints. GISA records only first construction, so demolition, rebuilding and urban greening are not shown, and buildings after 2021 are not there yet. Night lights are illustrative too.
- Address search uses OpenStreetMap's Nominatim. House numbers in Taiwan are incomplete, so when one is not found the search falls back to the lane, road section or district and says how precise the result is.
- Site stories (links with
?story=) average the same tree-cover values over a site outline from OpenStreetMap and over a ring around it; construction dates in the text come from the cited news and company sources. Fab buildings are real OpenStreetMap footprints (not illustrative); heights are mostly estimates, and the year each appears comes from GISA's first-built year or the year the site's tree cover went. Translucent ones are tagged as under construction in OpenStreetMap. While a story is open, no illustrative buildings are drawn inside the site.
- Revenue estimates and unicorns: TSMC reports revenue only for the whole company, not per fab. A story's revenue is an estimate: each month's reported revenue is split by that quarter's share of wafer revenue per process node and by the site's estimated share of each node's capacity (packaging sites: by the advanced-packaging share and each site's estimated part of it), converted to US$ at the year's average rate. The range reflects how certain each split is (about ±10–45 %). TSMC stopped publishing per-fab capacity after 2014, so later splits rely mostly on press and analyst estimates. Every US$1 bn of the estimate sends one unicorn out of the fab (just for fun); "trees per 🦄" converts the canopy area lost at a rough 50 m² of crown per tree. Inputs and sources:
pipeline/analysis/tsmc_revenue_inputs.json.
Sources
- Landsat Collection 2 Level-2 courtesy of the U.S. Geological Survey (processed in Google Earth Engine)
- ESA WorldCover 2021 v200, CC BY 4.0
- Year first built: GISA 1972–2021 (Ren, Huang, Yang & Zhou 2025, ISPRS J. Photogramm. Remote Sens. 220), CC BY 4.0
- Building height and built-up share: JRC GHSL P2023A (GHS-BUILT-H 2018, GHS-BUILT-S 2018 10 m), European Commission, Joint Research Centre
- Copernicus DEM GLO-30 © DLR e.V. 2010–2014 and © Airbus Defence and Space GmbH 2014–2018, provided under COPERNICUS by the European Union and ESA
- Address search and site outlines: Nominatim / © OpenStreetMap contributors (ODbL)
The data files (Apache Arrow) carry the UTM coordinates and source scene index of every pixel and open directly in Python (polars, pyarrow, DuckDB).